app.py

import gradio as gr
import torch
from transformers import pipeline
from timeit import default_timer as timer

username = "fmagot01"  ## Complete your username
model_id = f"{username}/distil-wav2vec2-finetuned-giga-speech"
device = "cuda:0" if torch.cuda.is_available() else "cpu"
pipe = pipeline("audio-classification", model=model_id, device=device)

def predict_trunc(filepath):
    preprocessed = pipe.preprocess(filepath)
    truncated = pipe.feature_extractor.pad(preprocessed,truncation=True, max_length = 16_000*30)
    model_outputs = pipe.forward(truncated)
    outputs = pipe.postprocess(model_outputs)

    return outputs


def classify_audio(filepath):
    """
      Goes from
      [{'score': 0.8339303731918335, 'label': 'Gaming'},
    {'score': 0.11914275586605072, 'label': 'Audiobook'},]
     to
     {"Gaming":  0.8339303731918335, "Audiobook":0.11914275586605072}
    """
    start_time = timer()
    #preds = pipe(filepath)
    
    preds = predict_trunc(filepath)
    
    outputs = {}
    pred_time = round(timer() - start_time, 5)
    for p in preds:
        outputs[p["label"]] = p["score"]
    return outputs, pred_time
    #return outputs


title = "Classifier of Audio Files"
description = """
This demo shows the application of the [distil-wav2vec2](https://huggingface.co/OthmaneJ/distil-wav2vec2) model fine tuned to the [gigaspeech](https://huggingface.co/datasets/speechcolab/gigaspeech) dataset. It will classify the audio provided to the domain of the content in it.
"""

filenames = ["audiobook.wav", "arts.wav", "news.wav"]
filenames = [[f"./{f}"] for f in filenames]
demo = gr.Interface(
    fn=classify_audio,
    inputs=gr.Audio(type="filepath"),
    outputs=[gr.outputs.Label(label="Predictions"),
             gr.Number(label="Prediction time (s)")
            ],
    title=title,
    description=description,
    examples=filenames,
)
demo.launch()